Fixes an off-by-one error when applying logits processor.

In this example, a draft model isn't used. The sample_idx is set to one less than the length of the input tokens, resulting in the last token being truncated.

(Pdb) idx
8
(Pdb) self._input_ids
array([    1,  1824,   349,   272,  1141,   302,  5459,  9955, 28804],
      dtype=int32)
(Pdb) len(self._input_ids)
9
(Pdb) self._input_ids[:idx]
array([   1, 1824,  349,  272, 1141,  302, 5459, 9955], dtype=int32)

@abetlen Would appreciate a review since you have a better understanding of the sample_idx argument. Especially considering this will impact logits = self._scores[idx, :]

Related: dottxt-ai/outlines#638